Neighborhood Sampling with Incremental Learning for Dynamic Network Embedding

Learning the network embeddings incrementally is attracting significant attention due to the dynamic and stable evolutionary properties exhibited by real-world networks. Existing methods update node embeddings by selecting the affected nodes, encompassing both the directly and indirectly affected nodes belonging to their neighboring nodes. However, most of these approaches select all the neighbors at each hop as indirectly affected nodes, overlooking the similarity differences between nodes due to different local structure and the diminishing impact as the neighborhood hops increases. To address these challenges, a new model called Neighborhood Sampling with Incremental Learning for Dynamic Network Embedding (NSIL-DNE) is proposed. Specifically, we employ a node similarity metric to measure the influence on the neighbors, and establish a sampling ratio to effectively select the closely topologically related neighbors as indirectly affected nodes. Extensive experiments conducted on real datasets show that NSIL-DNE outperforms comparison methods. Additionally, the comparison of sampling strategies experiment confirms the efficacy of the node similarity approach. And parameter analyses demonstrate that the decay property of the influence with increasing number of the neighborhood hops and not all the neighboring nodes at the same hop are affected equally.